TL;DR
- Anthropic publishes no token or dollar limit for Claude Pro, Max or Team. The official facts are multipliers: Max 5x is 5x Pro, Max 20x is 20x Pro, a Team Premium seat is 5x a Standard seat, plus weekly caps with no published number.
- Real usage is far above the sticker price. One log-based study hit the Max 20x weekly cap and measured about $1,860 of API-equivalent usage per week (about 1.9 billion tokens, mostly cache reads) on a $200/month plan. That is roughly 40x what the same spend would cost at list API prices.
- Max 5x was measured at at least $1,217 per week (cap never hit, so a lower bound). A single 5x session ran about 39M tokens before the limit stopped it.
- Per-model subscription limits are not published at all. The model tables below are our extrapolation from list prices, and are labeled that way.
What Anthropic actually publishes
The pricing page says plainly that there is "no fixed message count". Every plan has a rolling 5-hour session limit, and paid plans add weekly limits.
- Pro ($20/month, or $17 billed annually): Opus, Sonnet and Haiku. Fable 5.1 only through usage credits at API rates.
- Max 5x / 20x (from $100 and $200/month): 5x or 20x Pro usage per session. Fable included up to 50% of the weekly limit.
- Team: Standard seat $25/month ($20 annual), Premium seat $125/month ($100 annual). Premium is 5x Standard and includes Fable at 50% of weekly limits. Limits are per seat, not pooled.
Anthropic does not say how much more of your limit an Opus 5.5 turn uses than a Sonnet 5.5 turn.
What was actually measured
The best public data point we found is a September-October 2026 study that priced a developer's real Claude Code logs at the API list price in force each day and recorded when the plan limit stopped them. It is one user's workload, not a guarantee.
| Plan | API-equivalent per week | Tokens per week | Basis |
|---|---|---|---|
| Max 20x ($200) | about $1,860 | about 1.9B (range 1.8-2.3B), about 11M output | Measured, weekly cap hit twice |
| Max 5x ($100) | at least $1,217 | at least 1.3B, at least 7M output | Lower bound, cap never hit |
| Pro ($20) | not measured | not measured | Weekly cap not published |
Per 5-hour session: Max 5x stopped at a median of about 39M tokens (about $37 of API usage), observed twelve times in September 2026. Pro (about 8M) and Max 20x (about 155M) are projections from the multipliers, not measurements.
About 96% of those tokens were cache reads, the agent re-reading its own context. Cache reads are cheap per token, which is why a few billion tokens a week is possible. In the one model breakdown available (Opus 5, August 2026: 7.0B cache-read, 244M cache-write, 41.5M output, 0.2M fresh input), the cost was roughly $7,000 across the month at list prices.
Third-party estimates agree on the order of magnitude: about 225 Claude Code messages per window on Max 5x and about 900 on Max 20x, and one 38-day ccusage run on the $200 plan priced at about $3,200 at list prices. The limits have also moved: the 5-hour limit doubled on May 6, 2026, and was raised again around Sept 22.
Our own data: Hypercho's Claude Code logs
We counted the token usage recorded in our local Claude Code logs (token counts only, no message content) from July 2 to October 7, 2026, about 14 weeks, and priced it at the official list prices above.
| Model | Total tokens | Output tokens | API-equivalent |
|---|---|---|---|
| Opus | 64.0B | 74.8M | about $24,700 |
| Fable | 3.6B | 6.3M | about $2,600 |
| Sonnet | 3.2B | 2.9M | about $750 |
| Haiku | 26M | under 0.1M | under $1 |
| Total | 70.8B | 84M | about $28,000 |
- 333,000 model requests in 14 weeks, with 96% of tokens being cache reads. The agent mostly re-reads its own context, which is why the token count is so large and the dollar cost so low per token.
- Our two heaviest weeks hit about $6,700 and $7,500 of API-equivalent usage (16-18B tokens each, in the weeks of Sept 21 and Sept 28).
- If all of this ran on a single $200/month Max 20x seat, 14 weeks would cost about $650 in subscription fees, so we would have received roughly 43x that in API value. Our peak week would be about 160x the weekly share of the subscription price.
How to read it. Our workload is the best case for a flat-rate plan: long-running agent sessions that reuse a huge cached context and write very little. Light chat use will see far less. The figures are approximate: we priced every Opus request at the Opus 5.5 rates, older models cost more per token, and the logs cannot tell us how many accounts or seats produced the usage. Our peak weeks are also well above the $1,860 per week measured in the study above, so treat the multiple as a measurement of our workload and not a promise of what one seat delivers.
Official API prices (per million tokens)
| Model | Input | Output | Cache read | Cache write (5m) |
|---|---|---|---|---|
| Fable 5.1 | $10 | $50 | $0.25 | $12.5 |
| Opus 5.5 | $4 | $20 | $0.2 | $5 |
| Sonnet 5.5 | $2 | $10 | $0.1 | $2.5 |
| Haiku 5.5 (≤100k prompts) | $0.1 | $0.5 | $0.01 | $0.125 |
Haiku 5.5 rises to $0.50 input / $2.50 output for prompts over 100k tokens. Sonnet 5.5's $2/$10 is now the standard price. Opus 5.5 and Sonnet 5.5 cache hits cost 0.05x input, Fable 5.1 cache hits 0.025x. Batch API is 50% off.
Extrapolated tokens per week by model
This is our estimate, not a published or measured limit. We take the measured weekly dollar value and ask how many tokens that budget covers on each model, using a cache-heavy agent mix (90% cache reads, 6% cache writes, 2% input, 2% output; our assumption). It assumes Anthropic's internal metering tracks list price, which it may not. Anthropic says its limits weight pricier models more heavily, and Fable and Opus may count for more than this implies.
| Model | Max 20x (measured) $1,860/week | Max 5x (lower bound) $1,217/week |
|---|---|---|
| Fable 5.1 | 855M | 560M |
| Opus 5.5 | 1.9B | 1.3B |
| Sonnet 5.5 | 3.9B | 2.5B |
| Haiku 5.5 (≤100k prompts) | 65.3B | 42.7B |
Read the Haiku row as "very large", not as a promise. Treat each cell as an order of magnitude.
What about Pro and the $500 Team?
- Pro: no measured weekly figure. The session projection is about 8M tokens per 5-hour window, a fifth of Max 5x.
- Team at $500: we read this as four Premium seats at $125/month. A Standard seat is reported at roughly 1.25x Pro per session (third-party), so a Premium seat is about 6.25x Pro, slightly above Max 5x. Four of them are about 25x Pro in total, but limits are per person, so one heavy user is capped at a single seat's allowance. That is why Max 20x can beat a $500 Team plan for a single power user.
How to read this
- Max 20x is the best value by a wide margin for sustained agent work: the sticker price is only 10x Pro, for 20x the session usage.
- The subscription is not a dollar credit. Heavy cache-reading agent workloads get many times their sticker price in API value, and light chat use gets much less.
- Opus 5.5 cut list prices to $4/$20 from Opus 5's $5/$25. Fable 5.1 is 2.5x Opus 5.5, so route routine steps to Sonnet 5.5 or Haiku 5.5 and keep Fable for the hard steps.
- Claude 4.7 and later use a tokenizer that produces about 30% more tokens for the same text.
What to do with this at Hypercho
If you deploy an AI employee on a real workflow, price it from your own logs, not from headlines. Run the workflow, read your usage, then compare the plan against API cost for the model mix you need. Talk to us and we will design the model routing around your workflow.
Caveats
- The third-party measured figures come from one developer's Claude Code workload and are not a quantity you can count on every week.
- Max 5x is a lower bound, and Pro and Max 20x session figures are projections.
- No measured data exists for Sonnet 5.5, Haiku 5.5 or Fable 5.1 on subscriptions. The model table is extrapolated.
- Prices and limits were checked on October 8, 2026, and change often. Verify at claude.com/pricing and your own usage page.
Sources
- Anthropic API pricing docs
- Claude plans and pricing
- finopsllm: How many tokens do you get with Claude Pro and Max (measured, third party)
- Ten Invent: what $200 buys at API prices (third party)
- tokn.watch: Max 5x vs 20x (third party)
